VEFX-Reward as a training / selection signal — rebuttal experiments

Generated 2026-11-14. Reward scores are soft expected predictions; overall = GeoAgg with (IF,RQ,EE) weights (2,1,1), range 1–4.
Setup. 100 VEFX-Bench prompts (stratified, 11–12 per task category; seed 2026). Each source is cached as one 81-frame clip (~16 fps, 832×480 or 480×832) that is both the editor input and the “original video” for every judge. Each editor samples K=10 seeds (0–9) with identical settings. Selector: frozen VEFX-Reward-4B GeoAgg, ties → lower seed. Held-out evaluators: VEFX-Reward-32B, and a zero-shot rubric-prompted Qwen3-VL-32B-Instruct (not trained on VEFX data; the paper’s VLM-as-judge rubric). “Random” = expected score of one uniformly drawn candidate (mean over 10 seeds) — the K=1 baseline; “Seed-0” is one realised draw. Best-of-k for k<10 is the exact expectation over random k-subsets. CIs: 95% prompt-level bootstrap.
Editors: Kiwi-Edit (linyq/kiwi-edit-5b-instruct-only, 50 steps) · VACE (Wan2.1-VACE-1.3B, all-white mask + source as VACE context, original instruction as prompt, 50 steps, CFG 5, shift 16 — the VEFX-Bench VACE protocol on the 1.3B model).

Gain of Best-of-k over a single random sample

Δ GeoAgg of the 4B-selected candidate vs. a single random draw (k=1), under each evaluator; shaded = 95% paired bootstrap CI over prompts. Hover for values.

Per-dimension and per-category effect (held-out VEFX-Reward-32B)

Gallery

Editor